AI Director's Assistant: Plan Scenes and Story Beats

Why Planning, Not Generation, Is Now the Bottleneck

AI video generation crossed a threshold. Ask a model for a woman walking through rain-soaked neon streets and you will get something usable in a couple of attempts. Ask it for a ninety-second story with a beginning, a turn, and an ending, and the results usually collapse: characters change faces, props teleport, the emotional arc flatlines somewhere around shot four.

The tool was never the bottleneck. The bottleneck is planning. Generation models are excellent at rendering a described moment and terrible at inventing structure, continuity, and intent on your behalf. That is why the most useful addition to an AI video pipeline is not another rendering engine — it is a planning layer that behaves like a director's assistant: someone who reads your premise, breaks it into beats, proposes shots, tracks what each character is wearing, and warns you when sequence three contradicts sequence one.

This guide walks through that workflow end to end. It covers what an AI planning assistant should actually do, how to build a story blueprint before you generate a single frame, how to design scenes with real cinematic language, how to hold character and location consistency across dozens of shots, how to choose between generation models for each shot type, and the mistakes that ruin otherwise competent AI videos.

What an AI Director's Assistant Should Actually Do

The phrase gets used loosely, so it helps to define the job. A genuine planning assistant sits upstream of the renderer and does four things well.

Concept development and pressure-testing

You bring a rough idea. The assistant returns a logline, a central conflict, a protagonist with a want and an obstacle, and a list of the questions your premise has not answered yet. Good output here looks like notes from a skeptical producer: who is the antagonist, what does the protagonist lose if they fail, and why does this story need to be visual rather than written?

Scene breakdown and shot lists

Once the story holds together, the assistant decomposes it. A 90-second piece might become six to eight sequences, each two to four shots, with a stated dramatic function for every sequence and a stated information job for every shot. This is where planning tools earn their keep, because a shot list written by hand for an AI project is tedious and a shot list omitted entirely is fatal.

Continuity tracking

Faces, wardrobe, hair, props, time of day, weather, injuries, and emotional state all need to persist across shots that may be generated days apart on different models. A planning layer keeps a ledger you can paste into every prompt.

Production constraint awareness

Different models handle different shot types. A planning assistant that knows the difference between a slow push-in on a face and a chaotic crowd scene will route each shot to a sensible approach instead of sending everything to the same default setting.

What a planning assistant should not do is pretend to have taste. It can tell you your ending is rushed; it cannot tell you whether the rush is deliberate. That judgment stays with you.

Build the Blueprint Before You Generate a Frame

Skipping this step is the single most common reason AI videos feel like unrelated clips stitched together. The blueprint does not need to be long. It needs to be specific.

Step 1: The one-page premise

Write, in under 200 words: protagonist, world, inciting incident, central obstacle, cost of failure, and final image. The final image matters more than beginners expect — if you know the last frame before you generate the first, every earlier decision has a target to aim at.

Step 2: Beat sheet into sequences

Convert the premise into eight to twelve beats. Then group beats into sequences of two to four beats each. Each sequence gets a one-line dramatic function: establish the drought, introduce the rival, reveal the betrayal. If a sequence has no function, cut it. If two sequences have the same function, merge them.

Step 3: Sequences into shot lists

For each sequence, list shots with three fields: framing, action, and continuity notes. A realistic entry looks like this:

• Shot 12 — Medium close-up. Mara lowers the lantern; her face is lit from below. Continuity: torn left sleeve, ash smudge on right cheek, hair tied back.
• Shot 13 — Wide. Lantern light reveals the collapsed bridge in the background. Continuity: no rain yet, ground dry, late afternoon.
• Shot 14 — Insert. Her hand tightens on the rope. Continuity: rope frayed near the knot, same sleeve tear visible at frame edge.

Three fields per shot is enough. Ten fields per shot means you will stop filling them in by shot six.

Step 4: A generation order that protects continuity

Generate all shots featuring a character in one session, in chronological story order, even if you will edit them out of order. Models drift subtly over time and across sessions; batching shots by character and wardrobe reduces visible seams.

Designing Scenes With Real Cinematic Language

A scene in AI video is not a prompt. It is a set of decisions about space, light, and where the audience looks. Four levers do most of the work.

Composition geometry

Decide where the subject sits in frame and what the empty space is doing. A character placed on the left third with dark space to the right reads as anxious or watched. Centered and symmetrical reads as formal, ceremonial, or trapped. These are cheap decisions that carry enormous meaning, and AI models follow them reliably when you state them explicitly — subject on the left third, negative space to camera right, no other figures in frame.

Lens, height, and movement

Lens language translates directly into prompts. A wide lens close to the subject exaggerates depth and makes spaces feel bigger. A long lens compresses background and isolates a face. Camera height carries status: low angles enlarge, high angles shrink, eye level equalizes. Movement should have a reason — a slow push-in raises tension, a lateral track follows a decision, a handheld drift signals unease. State one movement per shot. Two movements in one prompt usually produces neither.

Light and color as narrative devices

Assign each sequence a lighting logic and keep it consistent. Practical sources — lanterns, screens, firelight — give AI models concrete motivation and produce more believable falloff than abstract descriptions like moody lighting. Color temperature can track the emotional arc: warm and saturated in the opening, desaturated and cool after the turn, a single warm accent returning in the final shot to signal hope.

Blocking and eyeline

Eyeline direction is a continuity trap. If your protagonist looks frame-left in shot A and frame-right in shot B, the audience reads it as a jump even if they cannot name why. Record eyeline direction in your continuity ledger alongside wardrobe. Blocking matters too: if a character stands on the left of the room, keep them on the left of the room unless a shot deliberately crosses the line.

Coverage strategy

Three shots cover most dramatic moments: a wide that establishes geography, a medium that carries performance, and a detail that carries emotion. Generate the wide first, use its best frame as an image reference for the medium, and the medium as reference for the detail. This chaining approach produces far more consistency than three independent text prompts.

Holding Characters, Props, and Locations Consistent

Consistency is a systems problem, not a prompting trick. Three habits solve most of it.

Build a reference pack per character

Create four to six approved stills covering front, three-quarter, and profile views, in the same lighting where possible. Approve one hero image and treat it as canonical. Every subsequent shot of that character should reference the hero image, either as an image-to-video start frame or as a style reference, rather than relying on description alone.

Keep a continuity ledger

A simple table beats memory. Columns: character, wardrobe, hair, props, physical state, and location state. Update it after every approved shot. When a prompt contradicts the ledger, the ledger wins — regenerating one shot is cheaper than reshooting six.

Use prompt scaffolding, not prose

Long descriptive paragraphs drift. Structured scaffolding repeats reliably:

[subject: Mara, 30s, dark braid, torn left sleeve, ash on right cheek]
[action: lifts lantern, turns toward bridge]
[environment: dry riverbed, late afternoon, no rain]
[camera: medium close-up, eye level, slow push-in]
[light: warm practical from below]
[style: grounded fantasy, natural palette, 35mm grain]

Keep the subject, style, and environment blocks identical across an entire sequence. Change only action and camera. This single practice removes most flicker and identity drift.

Lock locations with a plate shot

Generate one clean, empty wide shot of each location and reuse it as the visual anchor. When later shots must match, reference the plate. Locations with recognizable architecture — a bridge, a doorway, a specific window — benefit most.

Turning Shots Into Sequences: Editing Logic for AI Footage

AI footage behaves differently from camera footage, and editing should adapt.

Match on action, not on composition

Cut in the middle of a movement rather than between two static poses. If a hand is reaching for a rope at the end of shot A, begin shot B mid-reach. This masks the small inconsistencies that AI generation inevitably introduces.

Hold shots slightly longer than instinct suggests

Generated clips tend to be visually dense and motion-heavy in the first second. Trimming in tight often produces a jarring, over-caffeinated feel. Let each shot breathe for an extra half second and cut on the motivated beat.

Use sound as the continuity glue

Ambient sound, a consistent room tone, and music that crosses cuts do more for perceived coherence than any visual fix. If two shots stubbornly refuse to sit together, try overlapping a single continuous ambience beneath both before regenerating anything.

Build a rough cut before perfecting any shot

Assemble with your best available takes even if some are weak. Structure problems only become visible once you can watch the whole piece. Fixing them before you have refined individual shots is far cheaper.

Choosing the Right Generation Model for Each Shot

No single model is best at everything, and shot-level routing beats tool loyalty.

Match model to shot type

• Dialogue and close-ups: prioritize models with strong facial stability and subtle motion. Test with a two-second clip before committing to a full sequence.
• Wide establishing shots: prioritize environmental detail and camera-move smoothness over facial fidelity.
• Action and crowds: prioritize motion coherence and accept reduced fine detail, which the cut will hide anyway.
• Inserts and details: prioritize texture and consistency with a reference image.

Criteria that actually matter in evaluation

1. Continuity retention with image references — this matters more than raw realism.
2. Duration per generation — longer native clips reduce stitching artifacts.
3. Camera-move obedience — does the model respect a stated push-in or dolly?
4. Iteration speed — a fast, slightly weaker model you can test five times beats a slow, brilliant one you test once.
5. Predictable failure modes — you want to know in advance that a model struggles with hands, or crowds, or text in frame.

When to switch mid-project

Switching models is worth it when a single shot type consistently fails after three attempts on the same model, or when you need a capability the current model lacks. It is not worth it for a single imperfect clip you can hide in a cut. Note the model and settings for every approved shot in your continuity ledger so you can reproduce the look later.

A Worked Example: 60-Second Fantasy Teaser

Here is how the workflow looks on a real project, compressed.

Beat sheet

1. Mara crosses a dry riverbed carrying a lantern.
2. She finds the bridge collapsed.
3. A stranger offers to guide her across.
4. Mid-crossing, the stranger asks for payment she does not have.
5. She cuts the rope and continues alone.
6. Final image: her lantern, still lit, on the far bank.

Sequence and shot list (abridged)

• Sequence A — Approach (shots 1–3): wide of the riverbed, medium tracking behind Mara, insert of boots on cracked earth.
• Sequence B — Discovery (shots 4–6): wide revealing the broken bridge, close-up of her face reacting, insert of her hand on the frayed rope.
• Sequence C — The stranger (shots 7–10): low-angle reveal, over-the-shoulder of the offer, two-shot negotiation, close-up of the stranger's outstretched hand.
• Sequence D — The choice (shots 11–14): medium of Mara deciding, insert of knife, wide of the rope snapping, final shot of the lantern alone.

Production notes

Generate the empty riverbed plate first. Then produce all Mara shots in one session using a single hero reference image and a fixed style block. Produce stranger shots separately with their own reference. Assemble a rough cut at 60 seconds before refining any single clip, cut on action at the rope snap, and lay a continuous wind ambience beneath the entire piece. Total: fourteen shots, two reference packs, one ledger.

Common Mistakes That Break AI Storytelling

1. Generating before structuring. If you cannot summarize your story in three sentences, the model cannot render it.
2. Describing style in every prompt but continuity in none. Style blocks stay fixed; continuity fields carry the changes.
3. Two camera moves per shot. Pick one.
4. Unmotivated lighting. Name a visible light source.
5. Ignoring eyeline direction. Left, right, up, down — record it and honor it.
6. Editing too tight. AI motion needs room; cut on beats, not on panic.
7. Chasing realism over readability. A slightly stylized look that holds consistency beats a photoreal look that flickers.
8. No rough cut until the end. Structure errors compound.
9. Regenerating everything instead of one variable. Change one field per attempt so you learn what caused the failure.
10. Skipping sound. Silence makes even good AI footage feel synthetic.

FAQ

Do I need a dedicated planning tool, or can I plan in a document?
A document works if you are disciplined about the ledger and scaffolding. A dedicated assistant helps when projects grow past roughly twenty shots, because consistency tracking and shot routing become hard to hold in your head.

How long should an AI-generated video be?
For a first project, target 45 to 90 seconds. Longer pieces are entirely possible, but every additional minute multiplies continuity surface area.

Why do faces drift even with a reference image?
Usually because the reference is used loosely, the style block changes between shots, or the character is small in frame in the reference. Use a large, well-lit hero image and keep the subject block identical.

Is it better to generate in story order or by shot type?
Group by character and wardrobe, then within that group, work in story order. Batching by character reduces drift; story order keeps emotional continuity legible.

How many attempts should a shot get before I change approach?
Three. If three attempts on the same model with one variable changed each time all fail, change the framing, the reference, or the model — not the prompt wording.

Can I mix live-action footage with AI shots?
Yes, and it often improves credibility. Match grain, color temperature, and lens feel in post; shoot your live material with the AI shots' palette in mind from the start.

What is the minimum viable planning document?
One page: premise, beat list, sequence functions, shot list with framing and action, and a continuity ledger. That is enough for most short projects.

How do I handle dialogue in AI video?
Generate silent performances and add voice separately. Lip-sync tools have improved but still fight with scene-level continuity, and a clean voice track paired with well-timed cuts usually reads better than imperfect synchronization.

The bottom line: treat your planning layer as the director and your generators as the crew. Build the blueprint, lock the references, keep the ledger, route each shot to the model that suits it, and assemble a rough cut early. Do that consistently and the difference shows — not in any single frame, but in whether the whole piece holds together as a story.